feat: Gemma 4 on-device inference via LiteRT-LM (Android)

README.md

@@ -89,6 +89,7 @@ All data stays on your device. You bring your own LLM provider.
 | Xiaomi MIMO | Anthropic-compatible | MiMo-7B-RL |
 | OpenRouter | OpenAI-compatible | Access multiple providers via one API |
 | Ollama | OpenAI-compatible (local) | Run models locally on your device |
+| Gemma (On-Device) | LiteRT-LM (on-device) | Run Gemma 4 fully on-device, no API key needed. Android only |
 
 ## Install
 - **iOS**: Download on the [App Store](https://apps.apple.com/app/memexai/id6760325170)
@@ -203,7 +204,7 @@ flutter run
 | Platforms | iOS, Android |
 | Database | Drift (SQLite) |
 | State Management | Provider + MVVM |
-| LLM Providers | Gemini, OpenAI, Claude, Bedrock, Kimi, Qwen, Doubao, GLM, MiniMax, MIMO, OpenRouter, Ollama |
+| LLM Providers | Gemini, OpenAI, Claude, Bedrock, Kimi, Qwen, Doubao, GLM, MiniMax, MIMO, OpenRouter, Ollama, Gemma (On-Device) |
 | Agent Framework | dart_agent_core |
 
 ### Project Structure

README_CN.md

@@ -89,6 +89,7 @@ Memex 的多 Agent 智能底座是一套完全开放的自定义 Agent 系统 
 | 小米 MIMO | Anthropic 兼容 | MiMo-7B-RL |
 | OpenRouter | OpenAI 兼容 | 通过一个 API 访问多个提供商 |
 | Ollama | OpenAI 兼容（本地） | 在本地设备上运行模型 |
+| Gemma（端侧） | LiteRT-LM（端侧推理） | 完全在设备上运行 Gemma 4，无需 API Key，仅支持 Android |
 
 ## 安装
 - **iOS**: 在 [App Store](https://apps.apple.com/app/memexai/id6760325170) 下载
@@ -203,7 +204,7 @@ flutter run
 | 平台 | iOS、Android |
 | 数据库 | Drift (SQLite) |
 | 状态管理 | Provider + MVVM |
-| LLM | Gemini、OpenAI、Claude、Bedrock、Kimi、通义千问、豆包、智谱 GLM、MiniMax、MIMO、OpenRouter、Ollama |
+| LLM | Gemini、OpenAI、Claude、Bedrock、Kimi、通义千问、豆包、智谱 GLM、MiniMax、MIMO、OpenRouter、Ollama、Gemma（端侧） |
 | Agent 框架 | dart_agent_core |
 
 ### 项目结构

android/app/build.gradle.kts

@@ -113,4 +113,10 @@ flutter {
 
 dependencies {
     implementation("com.google.mlkit:text-recognition-chinese:16.0.1")
+    // Official LiteRT-LM Kotlin API (replaces flutter_gemma)
+    implementation("com.google.ai.edge.litertlm:litertlm-android:latest.release")
+    // Coroutines for async inference
+    implementation("org.jetbrains.kotlinx:kotlinx-coroutines-android:1.8.1")
+    // OkHttp for model downloads (respects system VPN/proxy)
+    implementation("com.squareup.okhttp3:okhttp:4.12.0")
 }

android/app/proguard-rules.pro

@@ -1,3 +1,7 @@
 -dontwarn com.google.mlkit.vision.text.devanagari.**
 -dontwarn com.google.mlkit.vision.text.japanese.**
 -dontwarn com.google.mlkit.vision.text.korean.**
+
+# LiteRT-LM
+-keep class com.google.ai.edge.litertlm.** { *; }
+-dontwarn com.google.ai.edge.litertlm.**

android/app/src/main/AndroidManifest.xml

@@ -80,6 +80,10 @@
         <!-- Don't delete the meta-data below.
              This is used by the Flutter tool to generate GeneratedPluginRegistrant.java -->
         <meta-data android:name="flutterEmbedding" android:value="2" />
+
+        <!-- LiteRT-LM GPU acceleration (official requirement per https://ai.google.dev/edge/litert-lm/android) -->
+        <uses-native-library android:name="libvndksupport.so" android:required="false"/>
+        <uses-native-library android:name="libOpenCL.so" android:required="false"/>
     </application>
     <!-- Required to query activities that can process text, see:
          https://developer.android.com/training/package-visibility and